Research On Chinese Opinion Extraction Based On Deep Learning

As a popular machine learning model,deep learning has been paid more attention by many researchers in the field of Natural Language Processing in recent years.This paper aims at applying deep learning in opinion extraction of Chinese commentary and sentiment classification,and it has an important research significance and application value.In case of few and undisclosed data sets in the study of Chinese comments,we obtain the comments data from Taobao via the web crawler.Then,we will train word vector by using the Taobao comments data,Sogou news data and Baidu Wikipedia data separately.The manually tagged data sets will be used to train,verify and test the model.We proposed a deep learning model of Chinese aspect term extraction based on memory network(Me-Bi LSTM).The contribution of this model is the integration of memory network and bidirection long short-term memory.The purpose of the model is to extract the sentence patterns from the historical sentences and apply them to the unknown comment sentences.The model consists of Bidirectional Long Short-Term Memory Network(Bi LSTM),Memory Network(MeN)and Conditional Random Field Algorithm(CRF).In the model,the BiLSTM is regarded as a feature extractor to extract the sentence characteristics.As a memory,MeN is used to store the sentence patterns extracted by Bi LSTM.CRF is used to calculate the joint probability of the whole tag sequence under the condition of given sentence features and sentence characteristics.Finally,the results will be obtained using Viterbi algorithm.Compared with the other 7 benchmark models,the effectiveness of the proposed model is demonstrated.Based on Me-Bi LSTM,we extended it to sentiment classification of Chinese reviews.We proposed a deep learning model of sentence level Chinese sentiment classification based on memory network(Me-BiLSTM-C).In this model,the label layer of Me-Bi LSTM is replaced by classification layer.Compared with the other 2 benchmark models,the effectiveness of the proposed model is demonstrated.